Skip to content

Potential new bottlenecks due to input quant/reshape operators #188

@fpjentzsch

Description

@fpjentzsch

KWS:

Image

The new Thresholding (input quant) comes before the flatten (realized implicitly by the DWC), so its max PE is 1, leading to an interval of 490 cycles, which is the new bottleneck:

"Thresholding_rtl_0": 490,
"MVAU_hls_0": 392,
"MVAU_hls_1": 256,
"MVAU_hls_2": 256,
"MVAU_hls_3": 384,
"LabelSelect_hls_0": 13

bnn-pynq TFC:

Image

The Reshape (no-op) performs the flattening of the 28x28 input. Its PE is limited to 28, so its lowest possible interval is 28 cycles. With the current folding this is not a problem (bottleneck 64 cycles), but I can imagine cases where it could be.

"Reshape_rtl_0": 28,
"Thresholding_rtl_0": 28,
"MVAU_hls_0": 64,
"MVAU_hls_1": 64,
"MVAU_hls_2": 64,
"MVAU_hls_3": 8,
"LabelSelect_hls_0": 11

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions